| paper | d4-conf-ifaamas-2025-0342 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | unclear |
The paper’s numbered Lemmas 1 and 2 establish accuracy bounds, not computational results under the stated anchor rule. Algorithm 1 has a running-time analysis, but it is not a numbered theorem, lemma, corollary, or proposition, so there is no qualifying result for a continuous mirror to carry forward.
fails bit a — no named computational result to mirror
The proposed mass-selection surrogate covers the odds-based subset selector; it does not cover the paper’s conditional-accuracy objective, conformal coverage guarantees, or empirical claims.
There is no qualifying numbered computational anchor in the paper. Lemma 1 proves an accuracy lower bound: the joint success probability is at least \(1-\alpha\) times the probability that the product of the experts’ conditional odds for the true class exceeds \(1\). Lemma 2 gives a condition, \(\epsilon>1\) and \(\alpha=0\), under which the conformal-set bound is tighter than the corresponding full-label-space bound. Neither is a result that a problem is tractable, hard, or has a stated computational formulation. Algorithm 1 is analyzed in the text as taking \(O(hc)\) time, but that is not a numbered theorem, lemma, corollary, or proposition. The empirical accuracy results are not numbered computational results either. So, under your anchor rule, I cannot honestly present a paper-anchored mirror.
The strongest *prospective* positive lead is a continuous version of the subset-selection computation in Algorithm 1. It is not a qualifying anchor, but it preserves a recognizable part of the paper’s problem. Imagine a large panel of experts reviewing standardized cases. A type \(t\) records an expert’s relevant response behavior: their initial label \(p_t\) for the case and their confusion matrix \(C^t\). Let \(\mu_t\) be the fraction of the eligible panel with type \(t\), and let \(A\) be the conformal set. For \(y\in A\), restrict and renormalize the type’s confusion-matrix row to \(A\), obtaining \(q^A_{t,y}\). Define
\[ R_{t,y}=\frac{q^A_{t,y}(p_t)}{1-q^A_{t,y}(p_t)}. \]
The Continuous Conformal Expert-Mass Selection problem is: given \(A\), the rational masses \((\mu_t)\), and the type response data, choose a pseudo-label \(y^*\in A\) and selected masses \(z_t\in[0,\mu_t]\) maximizing
\[ \sum_t z_t\log R_{t,y^*}. \]
A solution is the chosen pseudo-label and selected mass for each type; selected experts must make their final predictions from \(A\). This is the mass version of the paper’s product-of-odds selection criterion: when many experts share a type, the discrete product becomes a weighted sum of log odds. The optimum selects all mass with \(R_{t,y^*}>1\) and none with \(R_{t,y^*}<1\), then chooses the best pseudo-label. Thus this particular mirror looks tractable in time polynomial in the number of types, \(|A|\), and the input bit length.
The regime would be a large, centralized pool of experts repeatedly reviewing standardized cases, with many experts sharing a response profile and relatively few distinct profiles. Mass means a share of that panel, not a share of images or labels. This is plausible as a high-multiplicity analogue of *selecting experts for a case*: it retains the conformal restriction, confusion-based assessment, and subset-selection objective. It covers the greedy selector and its odds-based surrogate, not the paper’s full conditional-accuracy optimization, conformal coverage guarantees, or experimental claims. The expected class is A for this surrogate; whether exact conditional-accuracy selection under pooled voting is tractable remains a separate question, as do correlated experts and rounding the selected masses back to an actual panel.
The weakest point is the regime. The paper studies a small set of experts per image, and their responses can differ by case; it does not establish that a large panel with a modest number of stable expert types is realistic for its setting. A reviewer might accept the proposed problem as a high-multiplicity version of its selector, but could reasonably regard the type-level panel as a new staffing model rather than the continuous form of the paper’s main optimization problem.
The strongest case against a mirror starts with the anchor rule: this paper has no numbered result about computational complexity. Lemma 1 gives an accuracy bound, and Lemma 2 compares accuracy bounds under an assumption. Neither classifies an algorithmic problem. Algorithm 1’s \(O(hc)\) running time is stated in the text, but it is not a theorem about a complexity question. So there is no qualifying computational result here for a continuous mirror to carry forward.
The proposed mass-selection problem is coherent, and a large panel of annotators with repeated response profiles is plausible. But the proposed objective continues the paper’s odds-based surrogate, not its stated goal of maximizing conditional accuracy under a combination rule. For each pseudo-label \(y\), the objective separates by type: select all mass when \(\log R_{t,y}>0\), none when it is negative, and any amount when it is zero. Trying each \(y\in A\) takes \(O(\tau |A|)\). That is a clean high-multiplicity restatement of an already simple greedy selector, rather than a new complexity question surfaced by continuization.
A closer mirror would optimize actual conditional accuracy for a mass-selected panel, with the paper’s combination rule represented explicitly. In an infinite pool of independent experts, type-level vote shares become deterministic; without a panel-wide capacity or effort constraint, selection remains a separate decision for each case. Adding such constraints could create a substantial continuous allocation problem, but those constraints would supply a new staffing problem rather than continue a computational result established in this paper. There is also a modeling tension: using confusion matrices keeps the type space compact but may erase case-specific response differences; defining types by full response behavior can make the support as large as the expert pool.
This makes the negative case strong on paper-anchoring and weaker on the universal claim. A large annotation workforce with stable repeated profiles could make the high-multiplicity setting sensible, and an exact mass-based accuracy problem might be worth studying if the authors recognize it as their problem. I would argue against treating this paper as a compelling ChoCo anchor, while conceding that the broader possibility is not ruled out.
The adversarial triple: the proponent anchors on up to three named results; the opponent sees that case and must defeat every anchor; the judge decides which case convinced it. These are the pipeline’s own outputs, generated by tools/triple_run.py — no human edited them. The paper’s own text is not reproduced here beyond the quoted statement above.